# LLM benchmark scores by model, since 2019

**Compare large language models on more than 80 benchmarks, from GPQA Diamond and SWE-bench Verified to FrontierMath and Humanity's Last Exam.** The table gathers every result published in Epoch AI's Benchmarking Hub, with each score in its own unit, who ran the evaluation and a link to the source. Use it to rank models, follow how scores rose with each release, or set lab and paper numbers besid…

LLM benchmark scores by model, since 2019 is an open, machine-readable dataset published on Mostly Right as immutable Parquet versions, with a documented column schema, a JSON query API and a direct download. Its first rows are free to read and it is open to AI agents.

## Facts

- Slug: `llm-benchmark-scores-by-model-since-2019`
- Canonical page: https://mostlyright.md/datasets/llm-benchmark-scores-by-model-since-2019
- Publisher: mostlyrightmd (https://mostlyright.md/publishers/mostlyrightmd)
- Published: 2026-10-02T10:15:04.269Z
- Tables: 1
- Formats: Apache Parquet (download), JSON (query and sample)
- Cost: the first 500 rows of each table are free to everyone; querying and downloading need a free account and count against a monthly free allowance unless the workspace is subscribed
- Metadata API: `GET https://mostlyright.md/api/v2/public/datasets/llm-benchmark-scores-by-model-since-2019`
- Topics: llm benchmarks, large language models, ai evaluation, leaderboards, epoch ai, frontier models, ai capabilities

## How to use it

1. Read a table's schema — no account needed.
2. Fetch its published sample rows — no account needed.
   The first 500 rows of every table, in stored order, are free to everyone, with or without an account. Without an account you get that sample and nothing else: no filter, sort, search, download or page past row 500.
3. Sign in at https://mostlyright.md/login, connect this dataset to a workspace on its page, and create an `mr_use_` key.
4. Query bounded JSON, or download the complete Parquet snapshot and read it locally.
   A free workspace may read 100,000 rows and run 1,000 queries a UTC month past the sample. The allowance is shared by everyone in the workspace, and a free workspace may have any number of members. Reads count against the workspace whose data or key they use. A subscribed workspace is never counted, for anyone working in it.

Agents can do all of this over MCP at `https://mostlyright.md/api/mcp` — searching, reading schemas and sampling rows work with no credential at all.

## Table: llm_benchmark_results

- Table id: `9687578e-158b-4a35-a894-0152d5fba679`
- Slug: `llm-benchmark-results`
- Immutable version: `dde2b5b0-4267-5ba8-80d1-91188ea9f96c`
- Schema: `GET https://mostlyright.md/api/v2/public/tables/9687578e-158b-4a35-a894-0152d5fba679/schema`
- Sample rows, no key: `GET https://mostlyright.md/api/v2/public/tables/9687578e-158b-4a35-a894-0152d5fba679/preview`
- Bounded JSON query: `POST https://mostlyright.md/api/v2/public/tables/9687578e-158b-4a35-a894-0152d5fba679/query` with `x-api-key: mr_use_…`
- Full Parquet snapshot: `GET https://mostlyright.md/api/v2/public/tables/9687578e-158b-4a35-a894-0152d5fba679/current` with `x-api-key: mr_use_…` (application/vnd.apache.parquet)

### Columns (22)

| Column | Type | Description |
| --- | --- | --- |
| `result_id` | string | Identifier of the result: Epoch AI's own record id where it publishes one, otherwise a hash of the file name and the row, stable while the row is unchanged. |
| `benchmark` | string | The benchmark, named as Epoch AI names it, for example GPQA diamond or SWE-bench Verified. |
| `benchmark_release_date` | date | The date the benchmark was published, from Epoch AI's benchmark list. |
| `model` | string | The model the result belongs to, as Epoch AI groups versions of one model, for example GPT-5 or Claude Opus 4.5. |
| `model_version` | string | The exact model version evaluated, usually the developer's API identifier with Epoch AI's suffix for the reasoning setting; empty for one-per-family rows from… |
| `organization` | string | The organisation that developed the model; joint developers are listed together. |
| `country` | string | The country of the developing organisation. |
| `model_access` | string | How the model is released, for example API access or Open weights (unrestricted), from Epoch AI's model list. |
| `model_release_date` | date | The date the model version was released. |
| `training_compute_flop` | number | Epoch AI's estimate of the compute used to train the model, in floating-point operations, where it publishes one. |
| `score` | number | The score as the source publishes it, in the unit named by score_unit. |
| `score_unit` | string | The unit of score: fraction, percent, points out of 10, Elo rating, rating, speedup factor, Brier score, index points (0-100) or US dollars. |
| `score_fraction` | number | The score as a fraction between 0 and 1, for benchmarks scored as a proportion; empty for ratings, money and other units. |
| `higher_is_better` | boolean | Whether a higher score is better on this benchmark; false only for error measures such as the Brier score. |
| `stderr` | number | The standard error the source reports for the score, in the same unit as score. |
| `result_date` | date | The date the source attaches to the result: the run date for Epoch AI's own runs, otherwise the evaluation, grading, addition or last-update date the source st… |
| `reported_by` | string | Who ran or published the result: Epoch AI (its own run), Leaderboard (a benchmark or evaluator leaderboard), Model developer (the developer's own page), Paper… |
| `source` | string | The source Epoch AI cites for the result: a paper or leaderboard name, or the web host of its link. |
| `source_url` | string | The link to the cited source, or to Epoch AI's evaluation log for its own runs. |
| `evaluation_setup` | string | Settings that distinguish several results for one model on one benchmark, such as the number of shots, the agent harness or the reasoning effort. |
| `notes` | string | Epoch AI's note on the result, where it gives one. |
| `epoch_file` | string | The file in Epoch AI's download the row comes from. |

## Citing this dataset

Cite the dataset by name and canonical URL (https://mostlyright.md/datasets/llm-benchmark-scores-by-model-since-2019), the table, and the immutable `version_id` the rows came from. A query response carries `table.version_id` and `table.content_digest`; a download carries a strong ETag and a `Content-Digest` header.